Peking University releases OmniEdu, open education foundation models in 4B/9B/27B sizes
PekingUniversity · hf · 2026-09-22
Peking University introduces OmniEdu, an open family of foundation models for K-12 learning and teaching that spans problem solving, curriculum understanding, diagnosis, and tutoring rather than a single task.
- Training data combines 100+ educational resources organized around four capabilities: subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action/scaffolding, yielding 69,999 examples and 15.96M supervised tokens
- The pipeline uses deterministic cleaning, semantic auditing/rewriting, task-specific quality scoring, and token-budgeted diversity selection
- Fine-tuned at 4B, 9B, and 27B scales; OmniEdu-27B hits 63.12% EM / 76.69% F1 on K12-Bench, 85.89% on MathFish, 86.95% on EDUMATH, 78.74% on MathTutorBench Scaffold, and the top LongTutor Teaching average of 3.02
- Results show capability-balanced supervision effectively adapts general LLMs to educational tasks
More from Models
- Alibaba reported targeting 5-10T-parameter Qwen models, training Qwen 4 with 20 GW datacenter plan — mark_k · 2026-09-22
- EEBench comparison sparks debate: Grok 4.7 called out vs Astra's speed and cost — teortaxesTex · 2026-09-22
- Leak: OpenAI's new agent reportedly named Aeon, launch expected Thursday to rival Grok Bot — ZeroStateReflex · 2026-09-22
- Real telemetry contradicts Grok 4.7 ragebait: 46% fewer tokens per task — ns123abc · 2026-09-22
- Studying LLM psychology today is like psychology in 1850, researcher argues — repligate · 2026-09-22
- Meta's SAM 3.1 segmentation model spotted, used for GIF creation demos — Necessary-Garlic-704 · 2026-09-22